llama.cpp

c1b738ef - server : fix parallel generation with very small batch sizes